Embed Size (px)
DESCRIPTION
Presentation at Clusters, Clouds and Data for Scientific Computing 2014 Integrative analyses of large scale spatio-temporal datasets play increasingly important roles in many areas of science and engineering. Our recent work in this area is motivated by application scenarios involving complementary digital microscopy, radiology and “omic” analyses in cancer research. In these scenarios, the objective is to use a coordinated set of image analysis, feature extraction and machine learning methods to predict disease progression and to aid in targeting new therapies. I will describe tools and methods our group has developed for extraction, management, and analysis of features along with the systems software methods for optimizing execution on high end CPU/GPU platforms. Once having provided our current work as an introduction, I will then describe 1) related but much more ambitious exascale biomedical and non-biomedical use cases that also involve the complex interplay between multi-scale structure and molecular mechanism and 2) concepts and requirements for methods and tools that address these challenges.
Citation preview
Spa$o-‐temporal Sensor Integra$on, Analysis, Classifica$on
or
Can Exascale Cure Cancer?
EXASCALE CHALLENGES IN INTEGRATIVE MULTI-‐SCALE SPATIO-‐TEMPORAL ANALYSIS
Joel Saltz Chair Biomedical InformaIcs Stony Brook University
“Domain”: Spa$o-‐temporal Sensor Integra$on, Analysis, Classifica$on Big Data Extreme Compu$ng 2014
• Mul$-‐scale material/$ssue structural, molecular, func$onal characteriza$on. Design of materials with specific structural, energy storage proper$es, brain, regenera$ve medicine, cancer
• Integra$ve mul$-‐scale analyses of the earth, oceans, atmosphere, ci$es, vegeta$on etc – cameras and sensors on satellites, aircraO, drones, land vehicles, sta$onary cameras
• Digital astronomy • Hydrocarbon explora$on, exploita$on, pollu$on remedia$on
• Aerospace – wind tunnels, acquisi$on of data during flight
• Solid prin$ng integra$ve data analyses • Autonomous vehicles, e.g. self driving cars • Data generated by numerical simula$on codes – PDEs, par$cle methods
Typical Computa$onal/Analysis Tasks Spa$o-‐temporal Sensor Integra$on, Analysis, Classifica$on
• Data Cleaning and Low Level Transforma$ons • Data SubseWng, Filtering, Subsampling • Spa$o-‐temporal Mapping and Registra$on • Object Segmenta$on • Feature Extrac$on • Object/Region/Feature Classifica$on • Spa$o-‐temporal Aggrega$on • Diffeomorphism type mapping methods (e.g. op$mal mass transport)
• Par$cle filtering/predic$on • Change Detec$on, Comparison, and Quan$fica$on
Integra$ve Analysis: OSU BISTI NBIB Center
Big Data (2005) Associate genotype with phenotype Big science experiments on cancer,
heart disease, pathogen host response Tissue specimen -‐-‐ 1 cm3
0.1 μ resolu$on – roughly 1015 bytes
Molecular data (spa$al loca$on) can add addi$onal significant factor; e.g. 102 Mul$spectral imaging, laser
captured microdissec$on, Imaging Mass Spec, Mul$plex QD
Mul$ple $ssue specimens; another factor of 103
Total: 1020 bytes -‐-‐ 100 exabytes per big science experiment
The Tyranny of Scale (Oil Reservoir Management
Tinsley Oden - U Texas) process scale
field scale
km
cm
simulation scale
mm
pore scale
Why Applica$ons Get Big • Physical world or simula$on results • Detailed descrip$on of two, three (or more) dimensional space
• High resolu$on in each dimension, lots of $mesteps
• e.g. oil reservoir code -‐-‐ simulate 100 km by 100 km region to 1 km depth at resolu$on of 100 cm:
– 10^6*10^6*10^4 mesh points, 10^2 bytes per mesh point, 10^6 $mesteps -‐-‐-‐ 10^24 bytes (Yo1abyte) of data!!!
Center for Mul$ Scale Cancer Informa$cs (Sept 2014)
• Stony Brook • Oak Ridge Na$onal Labs • Emory • Yale
• Cancer Research meets HPC, Material Science, “omics”
• Vector Valued “omics”
Cen
ter
for
Com
preh
ensi
ve I
nfor
mat
ics
Integrative Cancer Research with Digital Pathology
histology neuroimaging
clincal\pathology
Integrated Analysis
molecular
High-resolution whole-slide microscopy
Multiplex IHC
Direct Study of Rela$onship Between vs
Lee Cooper, Carlos Moreno
Clustering identifies three morphological groups • Analyzed 200 million nuclei from 162 TCGA GBMs (462 slides)
• Named for functions of associated genes: Cell Cycle (CC), Chromatin Modification (CM), Protein Biosynthesis (PB)
• Prognostically-significant (logrank p=4.5e-4)
Feat
ure
Indi
ces
CC CM PB
10
20
30
40
500 500 1000 1500 2000 2500 3000
0
0.2
0.4
0.6
0.8
1
Days
Surv
ival
CCCMPB
VASARI Feature Set
Complex image analysis, feature extrac$on, machine learning pipelines SpaIo-‐temporal Sensor IntegraIon, Analysis,
ClassificaIon
Programming Tools • Mul$ level computa$onal pipeline management
• Region Templates – abstrac$on for mul$-‐scale spa$o-‐temporal computa$ons
• “Domain” specific language
Database Tools (Fusheng Wang) Spa$al Queries and Analy$cs
Introduction
• Feature based descriptive queries – Feature based filtering or feature aggregation – Spatial relationship based queries – Spatial join (two- or multi-), window, point-in-
polygon – Polygon overlay or spatial cross-matching – Distance based queries – Nearest neighbors
• Spatial analytics – Density based spatial patterns: find clusters,
hotspots, and anomalies – Spatial relationship modeling, e.g., geographically
weighted regression model(GWR)
Vector Valued “omics” Data Scale
Tumor Heterogeneity
Marusyk 2012
Whole Slide Imaging: Scale
Nature News and Comment - $1000 Genome
Brian Owens
Epigene$cs
Genetic Science Learning Center - Utah
University of Utah Epigenetics Training Site
Spanish National Cancer Research Center
EMR Data Analytics: Tools for Clinical Phenotyping and Population Health
Patient History
Physical Exams
Chemistries
Hematology
CT Scans
MRI Scans
PET Scans
Path specimens
Genomics
Proteomics
Metabalomics
Lipidomics
Integrative Predictive Analytics
Cen
ter
for
Com
preh
ensi
ve I
nfor
mat
ics
• Example Project: Find hot spots in readmissions within 30 days – What fraction of patients with a given principal diagnosis will
be readmitted within 30 days? – What fraction of patients with a given set of diseases will be
readmitted within 30 days? – How does severity and time course of co-morbidities affect
readmissions? – Geographic analyses
• Compare and contrast with UHC Clinical Data Base – Repeat analyses across all UHC hospitals – Are we performing the same? – How are UHC-curated groupings of patients (e.g., product
lines) useful?
Clinical Phenotype Characterization and the Emory Analytic Information Warehouse
Andrew Post, Sharath Cholleti, Doris Gao, Michel Monsour, Himanshu Rathod
Cen
ter
for
Com
preh
ensi
ve I
nfor
mat
ics
30-Day Readmission Rates for Derived Variables Emory Health Care
Cen
ter
for
Com
preh
ensi
ve I
nfor
mat
ics
Geographic Analyses UHC Medicine General Product Line (#15)
Analytic Information Warehouse
Cen
ter
for
Com
preh
ensi
ve I
nfor
mat
ics
Predictive Modeling for Readmission
• Random forests (ensemble of decision trees) – Create a decision tree using a random subset of the
variables in the dataset – Generate a large number of such trees – All trees vote to classify each test example in a
training dataset – Generate a patient-specific readmission risk for each
encounter • Rank the encounters by risk for a subsequent 30-
day readmission
Sharath Cholleti
Cen
ter
for
Com
preh
ensi
ve I
nfor
mat
ics
Predictive Modeling for 180 UHC Hospitals, 35 Million Patients Identify High Risk Patients! Readmission fraction of top 10% high risk patients
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1 9 17
25
33
41
49
57
65
73
81
89
97
105
113
121
129
137
145
153
161
169
177
185
All Hospital Model
Individual Hospital Model
DSRIP
Delivery System Reform Incen$ve Payment (DSRIP) Program
What is DSRIP?
• 8 billion dollar grant from CMS to NY State – 25% reduc$on over five years in avoidable hospitaliza$ons and ER visits in the Medicaid and uninsured popula$on
– Collabora$ve effort to implement innova$ve projects focused on
• System transforma$on • Clinical improvement • Popula$on health improvement
5 YEAR GOALS
• Create integrated care delivery system anchored by safety net providers
• Engage partners across the care delivery spectrum to create a county wide network of care
• AOer five years transi$on this network to an ACO which will contract with insurance providers on an at risk basis
The projects
• The chosen projects must address the most significant healthcare issues in the Suffolk County Medicaid and uninsured popula$on and address healthcare dispari$es—some examples – Behavioral health: BH and primary care integra$on – Adults: COPD, diabetes, HTN, renal failure – Children: Asthma – Hi risk OB/neonates—Esp. Hispanic and African American communi$es
DATA ANALYTICS
• County wide healthcare data will be collected • Near real-‐$me data analy$cs will be used to drive healthcare improvements – Analyzing success and failures to create fast turn-‐around improvement opportuni$es
– Analyze trends in disease and wellness popula$on wide
– Con$nuous analysis of outcomes • Testbed for Machine Learning
Conclusions
• Major applica$on areas – Exascale++ – Impact – “cure cancer”
• “Domains” – Spa$o-‐temporal Sensor Integra$on, Analysis, Classifica$on
– Integra$ve Predic$ve Analy$cs
• Agile extreme scale compu$ng
